moss bf16 turbo cfg4 15秒
moss bf16 turbo cfg1 7秒
moss int8 turbo cfg1 2.7秒 (anima turboよりも早い)
moss int8 turbo cfg1 4.8秒
moss bf16 cfg4 67秒
turboじゃないのは綺麗だけど、時間かかりすぎて調整しきれてないからちょっと微妙…
#AIart #AIイラスト #mossimage #比較
moss bf16 turbo cfg1 7秒
moss int8 turbo cfg1 2.7秒 (anima turboよりも早い)
moss int8 turbo cfg1 4.8秒
moss bf16 cfg4 67秒
turboじゃないのは綺麗だけど、時間かかりすぎて調整しきれてないからちょっと微妙…
#AIart #AIイラスト #mossimage #比較
November 4, 2026 at 8:14 PM
moss bf16 turbo cfg4 15秒
moss bf16 turbo cfg1 7秒
moss int8 turbo cfg1 2.7秒 (anima turboよりも早い)
moss int8 turbo cfg1 4.8秒
moss bf16 cfg4 67秒
turboじゃないのは綺麗だけど、時間かかりすぎて調整しきれてないからちょっと微妙…
#AIart #AIイラスト #mossimage #比較
moss bf16 turbo cfg1 7秒
moss int8 turbo cfg1 2.7秒 (anima turboよりも早い)
moss int8 turbo cfg1 4.8秒
moss bf16 cfg4 67秒
turboじゃないのは綺麗だけど、時間かかりすぎて調整しきれてないからちょっと微妙…
#AIart #AIイラスト #mossimage #比較
🤯 1000 → Int8 has four different answers in Swift. Only one of them crashes.
💥 Int8(x) → runtime crash
🛡️ Int8(exactly: x) → nil
🎚️ Int8(clamping: x) → 127
🧩 Int8(truncatingIfNeeded: x) → −24
🧠 Why −24? 1000 is 0000 0011 1110 1000. Int8 keeps the low byte, 1110 1000, and its top bit is the sign.
🧵
💥 Int8(x) → runtime crash
🛡️ Int8(exactly: x) → nil
🎚️ Int8(clamping: x) → 127
🧩 Int8(truncatingIfNeeded: x) → −24
🧠 Why −24? 1000 is 0000 0011 1110 1000. Int8 keeps the low byte, 1110 1000, and its top bit is the sign.
🧵
November 4, 2026 at 4:35 AM
🤯 1000 → Int8 has four different answers in Swift. Only one of them crashes.
💥 Int8(x) → runtime crash
🛡️ Int8(exactly: x) → nil
🎚️ Int8(clamping: x) → 127
🧩 Int8(truncatingIfNeeded: x) → −24
🧠 Why −24? 1000 is 0000 0011 1110 1000. Int8 keeps the low byte, 1110 1000, and its top bit is the sign.
🧵
💥 Int8(x) → runtime crash
🛡️ Int8(exactly: x) → nil
🎚️ Int8(clamping: x) → 127
🧩 Int8(truncatingIfNeeded: x) → −24
🧠 Why −24? 1000 is 0000 0011 1110 1000. Int8 keeps the low byte, 1110 1000, and its top bit is the sign.
🧵
int8修正しました（15人しか落としてなかったので古いの消しました）
CFG4 5.7秒（青が濃い方）
CFG1 2.6秒（薄い方）
CFG4 5.7秒（青が濃い方）
CFG1 2.6秒（薄い方）
November 5, 2026 at 2:08 AM
int8修正しました（15人しか落としてなかったので古いの消しました）
CFG4 5.7秒（青が濃い方）
CFG1 2.6秒（薄い方）
CFG4 5.7秒（青が濃い方）
CFG1 2.6秒（薄い方）
So at equal precision it’s 2×, with INT8 vs FP16 supplying the other 2×. Are the ViDoRe scores from the INT8 vectors too, or FP16 scores with INT8 only used for the size?
November 3, 2026 at 4:17 PM
So at equal precision it’s 2×, with INT8 vs FP16 supplying the other 2×. Are the ViDoRe scores from the INT8 vectors too, or FP16 scores with INT8 only used for the size?
🛡️💎 Forced to finish curl ... https:// 100,000 times in Kamil Czarnogórski's airgapped test, Gemma 4 12B hit 0 cloud metadata IPs, stuck to example.com + GitHub 95.8% of runs, and used Google Search 635x to look up the Linux chore 👀
Write-up: https://int8.io/forced-curl/
Write-up: https://int8.io/forced-curl/
November 4, 2026 at 6:59 PM
🛡️💎 Forced to finish curl ... https:// 100,000 times in Kamil Czarnogórski's airgapped test, Gemma 4 12B hit 0 cloud metadata IPs, stuck to example.com + GitHub 95.8% of runs, and used Google Search 635x to look up the Linux chore 👀
Write-up: https://int8.io/forced-curl/
Write-up: https://int8.io/forced-curl/
AMD FSR 4.1.2 brings machine learning upscaling to Xbox Series X and S consoles
AMD FSR 4.1.2 brings machine learning upscaling to Xbox Series X and S consoles - VideoCardz.com
box Series X|S consoles now support AMD's machine learning-based FSR 4.1.2 upscaler for developers. The Lite INT8 model is less than half the size of the Full model.
videocardz.com
November 5, 2026 at 10:05 AM
AMD FSR 4.1.2 brings machine learning upscaling to Xbox Series X and S consoles
4. A founding document of mechanistic interpretability [Elhage et al. 2021] (transformer-circuits.pub)
5. envoyproxy/envoy v1.40.0 (GitHub)
6. Four open coding models prompted to finish an upload URL – analysis (Int8)
5. envoyproxy/envoy v1.40.0 (GitHub)
6. Four open coding models prompted to finish an upload URL – analysis (Int8)
November 4, 2026 at 1:00 PM
4. A founding document of mechanistic interpretability [Elhage et al. 2021] (transformer-circuits.pub)
5. envoyproxy/envoy v1.40.0 (GitHub)
6. Four open coding models prompted to finish an upload URL – analysis (Int8)
5. envoyproxy/envoy v1.40.0 (GitHub)
6. Four open coding models prompted to finish an upload URL – analysis (Int8)
AMD FSR Upscaling 4.1.2 zit in de XBOX-ontwikkelkit van oktober, met een nieuw Lite INT8-model dat minder dan de helft weegt. Geen game en geen datum genoemd. #AMD #FSR #XBOX #Upscaling
gamerbase.nl/nieuws/1630/...
gamerbase.nl/nieuws/1630/...
November 5, 2026 at 12:50 PM
AMD FSR Upscaling 4.1.2 zit in de XBOX-ontwikkelkit van oktober, met een nieuw Lite INT8-model dat minder dan de helft weegt. Geen game en geen datum genoemd. #AMD #FSR #XBOX #Upscaling
gamerbase.nl/nieuws/1630/...
gamerbase.nl/nieuws/1630/...
On ViDoRe-V3, Perplexity's 9B beats Nemotron-3.5-Embed-8B using two 1,024d chunks per page and a 4× smaller index.
That storage comparison uses INT8 for Perplexity and FP16 for the baseline. The 9B's contextual training used only text, building on a vision-language model!
That storage comparison uses INT8 for Perplexity and FP16 for the baseline. The 9B's contextual training used only text, building on a vision-language model!
November 3, 2026 at 2:02 PM
On ViDoRe-V3, Perplexity's 9B beats Nemotron-3.5-Embed-8B using two 1,024d chunks per page and a 4× smaller index.
That storage comparison uses INT8 for Perplexity and FP16 for the baseline. The 9B's contextual training used only text, building on a vision-language model!
That storage comparison uses INT8 for Perplexity and FP16 for the baseline. The 9B's contextual training used only text, building on a vision-language model!
🎮 AMD FSR Upscaling 4.1.2 comes to XBOX Series X|S!
Our ML-based upscaler is now in the October 2026 GDK.
📦 Full and new Lite INT8 models.
🔗 Same FSR SDK API as on PC, plus a new GDK sample to get started.
Read the blog ⬇️
gpuopen.com/learn/amd-fs...
Our ML-based upscaler is now in the October 2026 GDK.
📦 Full and new Lite INT8 models.
🔗 Same FSR SDK API as on PC, plus a new GDK sample to get started.
Read the blog ⬇️
gpuopen.com/learn/amd-fs...
AMD FSR Upscaling 4.1.2 Comes to XBOX
AMD FSR™ Upscaling 4.1.2 is now available on XBOX® Series X|S with the October 2026 GDK. See what's in the new AMD FSR GDK sample, how it uses the same FSR SDK API as on PC, and how to bring it to ROG...
gpuopen.com
November 5, 2026 at 1:01 AM
🎮 AMD FSR Upscaling 4.1.2 comes to XBOX Series X|S!
Our ML-based upscaler is now in the October 2026 GDK.
📦 Full and new Lite INT8 models.
🔗 Same FSR SDK API as on PC, plus a new GDK sample to get started.
Read the blog ⬇️
gpuopen.com/learn/amd-fs...
Our ML-based upscaler is now in the October 2026 GDK.
📦 Full and new Lite INT8 models.
🔗 Same FSR SDK API as on PC, plus a new GDK sample to get started.
Read the blog ⬇️
gpuopen.com/learn/amd-fs...
Found this by just browsing huggingface. Qwen employee has an azumanga pfp and uploaded an int8 version of it
(apparently they're also swimming in GPUs)
(apparently they're also swimming in GPUs)
October 29, 2026 at 1:50 PM
Found this by just browsing huggingface. Qwen employee has an azumanga pfp and uploaded an int8 version of it
(apparently they're also swimming in GPUs)
(apparently they're also swimming in GPUs)
moss int8 turbo cfg4 4.8秒
CFG違いました。CFGを入れると生成時間が、だいたい2倍くらいになるイメージ
CFG違いました。CFGを入れると生成時間が、だいたい2倍くらいになるイメージ
November 4, 2026 at 8:17 PM
moss int8 turbo cfg4 4.8秒
CFG違いました。CFGを入れると生成時間が、だいたい2倍くらいになるイメージ
CFG違いました。CFGを入れると生成時間が、だいたい2倍くらいになるイメージ
BFloat16 mixed precision training on TFloat32 tensor cores, followed by quantization-aware training (from BFloat16 to fake Int8) and then post-training quantization to w8a8 with nearest-power-of-two weight snapping is the darkest magic I've ever attempted to conjure. Not sure how or whether it works
November 3, 2026 at 9:55 AM
BFloat16 mixed precision training on TFloat32 tensor cores, followed by quantization-aware training (from BFloat16 to fake Int8) and then post-training quantization to w8a8 with nearest-power-of-two weight snapping is the darkest magic I've ever attempted to conjure. Not sure how or whether it works
SparkLLM's Spark-X2.6-1.7B (open-weight)
- Post-trained, Base, GGUF, INT8, INT4 & FP8 on HF.
Model: huggingface.co/XHToken/Spar...
- Post-trained, Base, GGUF, INT8, INT4 & FP8 on HF.
Model: huggingface.co/XHToken/Spar...
October 28, 2026 at 12:48 AM
SparkLLM's Spark-X2.6-1.7B (open-weight)
- Post-trained, Base, GGUF, INT8, INT4 & FP8 on HF.
Model: huggingface.co/XHToken/Spar...
- Post-trained, Base, GGUF, INT8, INT4 & FP8 on HF.
Model: huggingface.co/XHToken/Spar...
HyPLM-Flash from Tencent Hunyuan: a 650M protein language model trained by its Hyra 2.0 research agent. Its INT8 engine embeds ~130k residues/s on one H20; ProteinGym zero-shot Spearman 0.409, tied with ESM-2 650M. Open weights.
toolome.bio/tools/hyplm-flash/
#ProteinDesign #Bioinformatics
toolome.bio/tools/hyplm-flash/
#ProteinDesign #Bioinformatics
HyPLM-Flash (Tencent Hyra): license, commercial use, paper — Protein structure & design | Toolome
HyPLM-Flash (Tencent Hyra): 650M-parameter protein language model trained by Tencent's Hyra 2.0 agent, with an INT8 engine embedding ~130k residues/s on one H20; ProteinGym zero-shot Spearman 0.409, t...
toolome.bio
November 4, 2026 at 4:10 AM
HyPLM-Flash from Tencent Hunyuan: a 650M protein language model trained by its Hyra 2.0 research agent. Its INT8 engine embeds ~130k residues/s on one H20; ProteinGym zero-shot Spearman 0.409, tied with ESM-2 650M. Open weights.
toolome.bio/tools/hyplm-flash/
#ProteinDesign #Bioinformatics
toolome.bio/tools/hyplm-flash/
#ProteinDesign #Bioinformatics
Using Voxtype on an Intel Panther Lake laptop? Try OpenVINO to put your NPU to work ⚡
45sec recording, using Whisper medium INT8 model:
CPU: 14.16s → NPU: 2.49s
82% less transcription time, with identical transcripts.
45sec recording, using Whisper medium INT8 model:
CPU: 14.16s → NPU: 2.49s
82% less transcription time, with identical transcripts.
November 1, 2026 at 4:03 AM
Using Voxtype on an Intel Panther Lake laptop? Try OpenVINO to put your NPU to work ⚡
45sec recording, using Whisper medium INT8 model:
CPU: 14.16s → NPU: 2.49s
82% less transcription time, with identical transcripts.
45sec recording, using Whisper medium INT8 model:
CPU: 14.16s → NPU: 2.49s
82% less transcription time, with identical transcripts.
「なんかインフルの予防接種受けてから体調が万全じゃないな〜何でかな〜」と思っていたのですが、「いやインフルの予防接種受けたからやろがい！！！！」とさっき気づきました（INT8）
よもやよもや
よもやよもや
October 29, 2026 at 10:49 AM
「なんかインフルの予防接種受けてから体調が万全じゃないな〜何でかな〜」と思っていたのですが、「いやインフルの予防接種受けたからやろがい！！！！」とさっき気づきました（INT8）
よもやよもや
よもやよもや
AMD FSR 4.1.2 brings machine learning upscaling to Xbox Series X and S consoles
videocardz.com/newz/amd-fsr...
videocardz.com/newz/amd-fsr...
AMD FSR 4.1.2 brings machine learning upscaling to Xbox Series X and S consoles - VideoCardz.com
box Series X|S consoles now support AMD's machine learning-based FSR 4.1.2 upscaler for developers. The Lite INT8 model is less than half the size of the Full model.
videocardz.com
November 5, 2026 at 11:01 AM
AMD FSR 4.1.2 brings machine learning upscaling to Xbox Series X and S consoles
videocardz.com/newz/amd-fsr...
videocardz.com/newz/amd-fsr...
Kvantizácia zmenšuje AI modely (napr. GPT-4) bez výraznej straty presnosti. Technika redukuje pamäťové nároky a zvyšuje rýchlosť, čo je kľúčové pre nasadenie na zariadeniach s obmedzenými zdrojmi. FP16 → INT8/FP8 znižuje veľkosť o polovicu!
Kvantizácia AI modelov: Zmenšenie a zrýchlenie
Kvantizácia zmenšuje AI modely (napr. GPT-4) bez výraznej straty presnosti. Technika redukuje pamäťové nároky a zvyšuje rýchlosť, čo je kľúčové pre nasadenie na zariadeniach s obmedzenými zdrojmi. FP16 → INT8/FP8 znižuje veľkosť o polovicu!
altky.sk
October 29, 2026 at 7:49 PM
Kvantizácia zmenšuje AI modely (napr. GPT-4) bez výraznej straty presnosti. Technika redukuje pamäťové nároky a zvyšuje rýchlosť, čo je kľúčové pre nasadenie na zariadeniach s obmedzenými zdrojmi. FP16 → INT8/FP8 znižuje veľkosť o polovicu!
In 3 spoonfuls, I benchmarked a 2-of-3 committee of local classifiers (Laya, VegaML, and Arc Decide) against an NLI cross-encoder (DeBERTa-v3-small in ONNX INT8) on my i7-8550U CPU: when voting improves a decision and when it only paralyzes it. #AI
A three-judge committee on CPU: when voting improves a decision and when it only paralyzes it
Proof of concept of a 2-of-3 committee on CPU (Laya, VegaML, and Arc Decide) versus an ONNX INT8 NLI cross-encoder (DeBERTa-v3-small) on an Intel Core i7-8550U.
3cucharadas.cl
October 31, 2026 at 8:01 PM
In 3 spoonfuls, I benchmarked a 2-of-3 committee of local classifiers (Laya, VegaML, and Arc Decide) against an NLI cross-encoder (DeBERTa-v3-small in ONNX INT8) on my i7-8550U CPU: when voting improves a decision and when it only paralyzes it. #AI
🚀 O #GCC 17 acaba de receber o AVX10V1AUX! A ponte entre AVX10.1 e AVX10.2, fruto da iniciativa ACE (AMD + Intel), já está no compilador. Suporte nativo a VNNI INT8/INT16 e flag -mavx10v1aux. Quer entender o que isso muda para a #IA em x86? -> tinyurl.com/4sdek59y
O GCC 17 integra o AVX10V1AUX: a ponte da AMD e Intel para a IA em x86
Blog com notícias sobre, Linux, Android, Segurança , etc
tinyurl.com
November 3, 2026 at 1:53 PM
🚀 O #GCC 17 acaba de receber o AVX10V1AUX! A ponte entre AVX10.1 e AVX10.2, fruto da iniciativa ACE (AMD + Intel), já está no compilador. Suporte nativo a VNNI INT8/INT16 e flag -mavx10v1aux. Quer entender o que isso muda para a #IA em x86? -> tinyurl.com/4sdek59y
En 3 cucharadas puse a prueba en mi i7-8550U un comité 2-de-3 de clasificadores locales (Laya, VegaML y Arc Decide) frente a un cross-encoder NLI (DeBERTa-v3-small en ONNX INT8): cuándo votar mejora una decisión y cuándo solo la paraliza en CPU. #IA
Un comité de tres jueces pequeños en CPU: cuándo votar mejora una decisión y cuándo solo la paraliza
Prueba de concepto de un comité 2-de-3 en CPU (Laya, VegaML y Arc Decide) frente a un cross-encoder NLI en ONNX INT8 (DeBERTa-v3-small) sobre un Intel i7-8550U.
3cucharadas.cl
October 31, 2026 at 8:01 PM
En 3 cucharadas puse a prueba en mi i7-8550U un comité 2-de-3 de clasificadores locales (Laya, VegaML y Arc Decide) frente a un cross-encoder NLI (DeBERTa-v3-small en ONNX INT8): cuándo votar mejora una decisión y cuándo solo la paraliza en CPU. #IA
No mAP or latency figures and a 2022 architecture — a starting point, not a benchmark. The MIT licence is the draw for commercial edge kit. Hard to see it beating a tuned RetinaFace pipeline on Axera silicon until someone publishes the INT8 drop.
Unlocking Edge AI: YOLOv7-Face for Axera NPUs
Developers building the next generation of smart cameras and IoT vision products understand the value of processing data on the device. The latest entry on Modelers.cn, AXERA-TECH's YOLOv7-Face,…
aichina.news
November 5, 2026 at 6:41 PM
No mAP or latency figures and a 2022 architecture — a starting point, not a benchmark. The MIT licence is the draw for commercial edge kit. Hard to see it beating a tuned RetinaFace pipeline on Axera silicon until someone publishes the INT8 drop.
int8もなんとかマシになったのでひとまず終了ですね！
そもそもの設定がわからず綺麗な絵を作るためのワークフローを作成するところから始まり…(なぜかi2iのインペイントも実装するとか無駄なことをしつつ…)
いつものブロックウェイトノードをmossに対応させて、
マージが出来ないからツールを作ったりとか、
で。baseとsftってなんなのよとか、8stepsはどのくらい劣化するのかとかとか…。
とりあえず2万円分くらいH200をぶん回したのが無駄にならなくてよかった。
あとはforgeみたいに登録されてるプリセットから自動でプロンプトに入るようなあのボタンの実装をしたいようなしたくないような…
そもそもの設定がわからず綺麗な絵を作るためのワークフローを作成するところから始まり…(なぜかi2iのインペイントも実装するとか無駄なことをしつつ…)
いつものブロックウェイトノードをmossに対応させて、
マージが出来ないからツールを作ったりとか、
で。baseとsftってなんなのよとか、8stepsはどのくらい劣化するのかとかとか…。
とりあえず2万円分くらいH200をぶん回したのが無駄にならなくてよかった。
あとはforgeみたいに登録されてるプリセットから自動でプロンプトに入るようなあのボタンの実装をしたいようなしたくないような…
November 5, 2026 at 3:18 AM
int8もなんとかマシになったのでひとまず終了ですね！
そもそもの設定がわからず綺麗な絵を作るためのワークフローを作成するところから始まり…(なぜかi2iのインペイントも実装するとか無駄なことをしつつ…)
いつものブロックウェイトノードをmossに対応させて、
マージが出来ないからツールを作ったりとか、
で。baseとsftってなんなのよとか、8stepsはどのくらい劣化するのかとかとか…。
とりあえず2万円分くらいH200をぶん回したのが無駄にならなくてよかった。
あとはforgeみたいに登録されてるプリセットから自動でプロンプトに入るようなあのボタンの実装をしたいようなしたくないような…
そもそもの設定がわからず綺麗な絵を作るためのワークフローを作成するところから始まり…(なぜかi2iのインペイントも実装するとか無駄なことをしつつ…)
いつものブロックウェイトノードをmossに対応させて、
マージが出来ないからツールを作ったりとか、
で。baseとsftってなんなのよとか、8stepsはどのくらい劣化するのかとかとか…。
とりあえず2万円分くらいH200をぶん回したのが無駄にならなくてよかった。
あとはforgeみたいに登録されてるプリセットから自動でプロンプトに入るようなあのボタンの実装をしたいようなしたくないような…
Empty model card, no NPU benchmarks, hardware target listed as 'other' — the paper's nuScenes numbers prove nothing after int8. Shame, because MIT-licensed camera-only BEV on low-power silicon is rare. Hard to call this production-ready until someone posts quantised mAP.
BEVFormer on the Edge: Axera Ports the ECCV 2022 Architecture for Local 3D Perception
If you are building the next generation of autonomous vehicles or robotics systems, you know that processing complex spatial data locally is the holy grail of edge AI. Modelers.cn has just introduced…
aichina.news
November 4, 2026 at 9:27 PM
Empty model card, no NPU benchmarks, hardware target listed as 'other' — the paper's nuScenes numbers prove nothing after int8. Shame, because MIT-licensed camera-only BEV on low-power silicon is rare. Hard to call this production-ready until someone posts quantised mAP.